Hacker News Reader

Comment by wrnr | Hacker News Reader

I think GPT-3 is a technical marvel but it did not reveal anything about language that wasn't already known by linguists for at least a 100 years. I've read this funny dialog generated by GPT-3 between Bezos and Page, but even that was edited by the "author". At the end of the day it nothing more than an advanced form of Dada poetry: https://en.wikipedia.org/wiki/Dada#Poetry

View on HN

It’s manipulative power is more what I’m interested in. I want to see what percentage of the general population can be manipulated by it. Could care less if it’s a pure design that simulates mastery of a language.

I have a hunch it will work on 50% of people. I’m willing to bet your average 13 year old would gobble up gpt articles without knowing any better.

I suspect very many adults would also gobble up GPT articles without knowing any better too. Just witness the production machine of most news organizations. They can write entire "stories" around a single tweet. The new data is less than 140 characters and that gets turned into a 5 minute spot.

To be fair, a lot of tweets don't make any sense out of context, so to make them comprehensible, you might need to add a lot of context.

Depends how you define 'manipulative power'. If you mean 'succeed in its human-defined objective of convincing people the output is generated by a human' I think it's evident that it is already often succeeding where other software fails. And certainly it can generate sentences based on any propaganda its curators care to stuff it with which are both reasonably coherent and sufficiently novel in terms of actual words used to pass plagiarism detectors. This would be a huge problem if people believed arguments stated by computers were more correct than those advanced by humans.

But we've got a lot more to fear from the sort of humans that would want to use GPT-3 to rehash their propaganda, because most of them are perfectly capable of stringing a sentence together themselves as well as defining their own goals and a sequence of actions to achieve that. And bribing, blackmailing and pointing guns

That depends the population's knowledge more than the machine itself used. According to accounts Christopher Columbus scammed surrender of natives through exploiting astronomy. If they had better local knowledge he would have been mocked at best and at worst would have become a lost expedition.

I'm actually quite worried about its abilities and effect it may have on the internet in more anonymous settings. It can easily trick humans not really paying attention and looking for issues.

How can you tell what is real and isn't without taking time to really look at everything thats written, more than just simply reading the text in full? I already have this problem, and GPT-3 honestly breaks the entire internet of discussion for me. It's great technology, but I worry we'll know even less is real.

Word salad

https://en.m.wikipedia.org/wiki/Word_salad

>I have a hunch it will work on 50% of people.

The number of people on Social Media that don't read past an article's headline is enormous, they are convinced by basic false headlines like "George Soros paying for mind control chips in Covid Vaccine" and it's astonishing. "I saw it on the internet so it must be true" used to be a joke when I was a kid and now its actually true for massive numbers of people.

>I think GPT-3 is a technical marvel but it did not reveal anything about language that wasn't already known by linguists for at least a 100 years.

Yet. It's so incredibly sensitive to input conditions I would say that we really don't know exactly how to explore it fully. I would also argue that we don't really understand a single language until we understand many, and until we tag and classify the inputs I would say we don't even really understand what we passed through the system at all.

Reply on news.ycombinator.com